celcilin/cellularflow-v4

🧬 CellularFlow v4: Memory-Augmented Continual Learning LLM

Author: Celcilin C S
Codebase: https://github.com/celcilin/cellularflow
Hugging Face: https://huggingface.co/celcilin/cellularflow-v4
License: MIT

CellularFlow is a memory-augmented neural architecture designed as a continual-learning alternative to standard Transformers. By replacing dense Feed-Forward Networks (FFN/MLP) with Multi-Head Associative DNA Memory Banks and an Episodic Memory Slot Buffer, CellularFlow decouples factual knowledge storage from sequence reasoning.

It achieves 83.9% domain retention across sequential domains (eliminating catastrophic forgetting) and enables zero-backprop streaming learning during inference.

---

🌟 Benchmark Results

Architecture	Parameters	Perplexity	Accuracy	5-Domain Retention
GPT-mini (Vanilla Transformer)	810K	8.51	36.4%	61.8%
CellularFlow v4 (Hybrid CMC)	379K (2.1× fewer)	2.54 (−70.3%)	73.7%	83.9% (+22.1 pp)

---

🚀 Quickstart: Loading from Hugging Face

1. Install CellularFlow

pip install git+https://github.com/celcilin/cellularflow.git

2. Run Inference

import torch
from huggingface_hub import hf_hub_download
from cellularflow import CellularFlowLM, CellularFlowTrainer

REPO_ID = "celcilin/cellularflow-v4"

# Download model weights and tokenizer from Hugging Face
weights_file = hf_hub_download(repo_id=REPO_ID, filename="CMC_BaseModel.pt")

# Instantiate CellularFlow v4
model = CellularFlowLM(
    vocab_size=8192,
    dim=512,
    n_layers=6,
    n_heads=8,
    n_entries=256,
    context_len=256,
    use_episodic=True
)

# Load pretrained weights
state_dict = torch.load(weights_file, map_location="cpu")
model.load_state_dict(state_dict)
model.eval()

print("CellularFlow v4 loaded successfully!")

---

🔬 Continual Learning Regimes

1. Mode 1 — Live Learn (inference_write=True): Updates DNA memory values on the fly via Exponential Moving Average (EMA) with zero backward pass. Protected by Spherical Anisotropy Regularization.
2. Mode 2 — Selective Fine-Tuning (set_mode("selective")): Freezes ~85% of the backbone and updates only DNA memory banks, eliminating catastrophic forgetting.
3. Mode 3 — Episodic Fact Injection (inject_fact()): Writes facts into slot-based episodic memory with temporal decay and post-epoch consolidation.

---

📖 Citation

@article{celcilin2026cellularflow,
  title={CellularFlow: A Memory-Augmented Continual-Learning Architecture Decoupling Associative Memory from Sequence Reasoning},
  author={Celcilin C S},
  journal={arXiv preprint},
  year={2026},
  url={https://github.com/celcilin/cellularflow}
}